POLED - Probabilistic Online Event Downsampling

Code for Probabilistic Online Event Downsampling, presented at CVPR Event Vision Wokshop 2025.

⭐ Best Paper Finalist Award ⭐

👉 Link to paper 👈

Configuration

First things first

Clone the repository

git clone https://github.com/DensoITLab/POLED

References

Export path variables (they are used in the scripts).

• $POLED_PATH --> Your root folder containing the code (usually .../POLED).
• $POLED_DATA --> Your folder containing all the datasets. As a guide, it will be set to $POLED_PATH/data, which contains some samples from NCaltech101.

export POLED_PATH=YOURPATH/POLED
export POLED_DATA=$POLED_PATH/data

1. Docker

If you use docker, congrats! You avoided yourself a lot of trouble. $POLED_PATH/docker_env/Dockerfile has some packages that are not needed to replicate the paper (e.g. torch, fiftyone), feel free to delete them, but remember to also erase the imports from the code.

To build the docker image just run:

cd $POLED_PATH
bash build_docker.sh

To start the docker container:

bash docker_env/run_docker.sh

2. No Docker

You can install all the packages needed with

cd $POLED_PATH
pip install -r requirements.txt

The bash scripts assume docker (e.g. see docker exec some_command in run_poled.sh). You can change these instances for your_python_environment or similar.

Quick Example (Downsampling only)

In a new terminal, re-export the variables if needed

export POLED_PATH=YOURPATH/POLED
export POLED_DATA=$POLED_PATH/data

and run the example

cd $POLED_PATH
bash scripts/run_experiments.sh

The downsampled data should be generated inside $POLED_DATA/NCaltech101/sampling for the downsamplers uniform and poisson at a downsampling rate of 0.05.

Custom datasets / methods

For custom testing on several datasets, first (1) generate the config files needed to match your folder structure and data paths. You will have to modify config/master.yaml, as well as config/datasets/datasets_paths.yaml.

Second (2), you will have to modify external/processor.py to prepare the dataloader.

Third (3), you will have to modify src/event_sampling.py to reference to the new data.

With this you should be able to downsample your data.

If you want to run the data in a method as well, you will have to make a script (e.g. scripts/run_NEWMETHOD.sh). Follow the existing scripts as a guide.

Each of the different methods tested in the paper requires its own Dockerfile. E.g. if you check docker_id inside config/methods/rvt.yaml you will see that it is point to the already running docker container for RVT.

Results

Citation

If you find this research useful, please consider citing it.

@InProceedings{Girbau-Xalabarder_2025_CVPR,
    author    = {Girbau-Xalabarder, Andreu and Nagata, Jun and Sumiyoshi, Shinichi},
    title     = {Probabilistic Online Event Downsampling},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2025},
    pages     = {4866-4874}
}